wellally.tech

  The democratization of medical AI is moving at a breakneck pace. With the release of open-weights models like Google’s Gemma and its fine-tuned medical variants, we are witnessing a shift from centralized, proprietary healthcare algorithms to distributed, accessible diagnostic tools. However, as we fine-tune these models on PubMed abstracts and clinical notes, we run into the classic alignment problem—but with higher stakes. In general-purpose LLMs, a hallucination is annoying. In MedGemma models, a hallucination or a bias in triage recommendation can be fatal. This brings us to a fascinating convergence of technologies: the specific, domain-expert architecture of models like MedGemma, and the alignment philosophies pioneered by Anthropic, specifically Constitutional AI . The Limits of RLHF in Medicine Traditionally, models are aligned using Reinforcement Learning from Human Feedback (RLHF). Human labelers rate two model outputs, and the model learns a reward function. In the me...